From 2c30f5052906b5e757943960df52879dcc3e37a3 Mon Sep 17 00:00:00 2001 From: eihrul Date: Sat, 11 Feb 2012 17:16:42 +0000 Subject: [PATCH 1/1] SIMD-ize the matrix multiplication in AnimateVertices git-svn-id: svn://svn.icculus.org/twilight/trunk/darkplaces@11678 d7cf8633-e32d-0410-b094-e92efae38249 --- mod_skeletal_animatevertices_sse.c | 150 ++++++++++++++++++++++++----- 1 file changed, 124 insertions(+), 26 deletions(-) diff --git a/mod_skeletal_animatevertices_sse.c b/mod_skeletal_animatevertices_sse.c index 044a250b..a25a9224 100644 --- a/mod_skeletal_animatevertices_sse.c +++ b/mod_skeletal_animatevertices_sse.c @@ -16,7 +16,6 @@ void Mod_Skeletal_AnimateVertices_SSE(const dp_model_t * RESTRICT model, const f matrix4x4_t *bonepose; matrix4x4_t *boneposerelative; float m[12]; - matrix4x4_t mm, mm2; const blendweights_t * RESTRICT weights; int num_vertices_minus_one; @@ -34,23 +33,74 @@ void Mod_Skeletal_AnimateVertices_SSE(const dp_model_t * RESTRICT model, const f { for (i = 0;i < model->num_bones;i++) { - // relativetransforms is in GL column-major order, which is what we need for SSE - // transposed style processing + const float * RESTRICT n = model->data_baseboneposeinverse + i * 12; + matrix4x4_t * RESTRICT s = &skeleton->relativetransforms[i]; + matrix4x4_t * RESTRICT b = &bonepose[i]; + matrix4x4_t * RESTRICT r = &boneposerelative[i]; + __m128 b0, b1, b2, b3, r0, r1, r2, r3, nr; if (model->data_bones[i].parent >= 0) - Matrix4x4_Concat(&bonepose[i], &bonepose[model->data_bones[i].parent], &skeleton->relativetransforms[i]); + { + const matrix4x4_t * RESTRICT p = &bonepose[model->data_bones[i].parent]; + __m128 s0 = _mm_loadu_ps(s->m[0]), s1 = _mm_loadu_ps(s->m[1]), s2 = _mm_loadu_ps(s->m[2]); +#ifdef OPENGLORIENTATION +#define SKELETON_MATRIX(r, c) _mm_shuffle_ps(s##c, s##c, _MM_SHUFFLE(r, r, r, r)) +#else +#define SKELETON_MATRIX(r, c) _mm_shuffle_ps(s##r, s##r, _MM_SHUFFLE(c, c, c, c)) +#endif + __m128 pr = _mm_load_ps(p->m[0]); + b0 = _mm_mul_ps(pr, SKELETON_MATRIX(0, 0)); + b1 = _mm_mul_ps(pr, SKELETON_MATRIX(0, 1)); + b2 = _mm_mul_ps(pr, SKELETON_MATRIX(0, 2)); + b3 = _mm_mul_ps(pr, SKELETON_MATRIX(0, 3)); + pr = _mm_load_ps(p->m[1]); + b0 = _mm_add_ps(b0, _mm_mul_ps(pr, SKELETON_MATRIX(1, 0))); + b1 = _mm_add_ps(b1, _mm_mul_ps(pr, SKELETON_MATRIX(1, 1))); + b2 = _mm_add_ps(b2, _mm_mul_ps(pr, SKELETON_MATRIX(1, 2))); + b3 = _mm_add_ps(b3, _mm_mul_ps(pr, SKELETON_MATRIX(1, 3))); + pr = _mm_load_ps(p->m[2]); + b0 = _mm_add_ps(b0, _mm_mul_ps(pr, SKELETON_MATRIX(2, 0))); + b1 = _mm_add_ps(b1, _mm_mul_ps(pr, SKELETON_MATRIX(2, 1))); + b2 = _mm_add_ps(b2, _mm_mul_ps(pr, SKELETON_MATRIX(2, 2))); + b3 = _mm_add_ps(b3, _mm_mul_ps(pr, SKELETON_MATRIX(2, 3))); + b3 = _mm_add_ps(b3, _mm_load_ps(p->m[3])); + } else - memcpy(&bonepose[i], &skeleton->relativetransforms[i], sizeof(matrix4x4_t)); - - // create a relative deformation matrix to describe displacement - // from the base mesh, which is used by the actual weighting - Matrix4x4_FromArray12FloatD3D(&mm, model->data_baseboneposeinverse + i * 12); // baseboneposeinverse is 4x3 row-major - Matrix4x4_Concat(&mm2, &bonepose[i], &mm); - Matrix4x4_Transpose(&boneposerelative[i], &mm2); // TODO: Eliminate this transpose + { + b0 = _mm_loadu_ps(s->m[0]); + b1 = _mm_loadu_ps(s->m[1]); + b2 = _mm_loadu_ps(s->m[2]); + b3 = _mm_loadu_ps(s->m[3]); +#ifndef OPENGLORIENTATION + _MM_TRANSPOSE4_PS(b0, b1, b2, b3); +#endif + } + _mm_store_ps(b->m[0], b0); + _mm_store_ps(b->m[1], b1); + _mm_store_ps(b->m[2], b2); + _mm_store_ps(b->m[3], b3); + nr = _mm_loadu_ps(n); + r0 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0))); + r1 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1))); + r2 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2))); + r3 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3))); + nr = _mm_loadu_ps(n+4); + r0 = _mm_add_ps(r0, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0)))); + r1 = _mm_add_ps(r1, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1)))); + r2 = _mm_add_ps(r2, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2)))); + r3 = _mm_add_ps(r3, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3)))); + nr = _mm_loadu_ps(n+8); + r0 = _mm_add_ps(r0, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0)))); + r1 = _mm_add_ps(r1, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1)))); + r2 = _mm_add_ps(r2, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2)))); + r3 = _mm_add_ps(r3, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3)))); + _mm_store_ps(r->m[0], r0); + _mm_store_ps(r->m[1], r1); + _mm_store_ps(r->m[2], r2); + _mm_store_ps(r->m[3], r3); } } else { - float originscale = -model->num_posescale; for (i = 0;i < model->num_bones;i++) { const short * RESTRICT pose7s = model->data_poses7s + 7 * (frameblend[0].subframe * model->num_bones + i); @@ -64,7 +114,7 @@ void Mod_Skeletal_AnimateVertices_SSE(const dp_model_t * RESTRICT model, const f dy = -tx*rz + ty*rw + tz*rx, dz = tx*ry - ty*rx + tz*rw, dw = -tx*rx - ty*ry - tz*rz, - scale; + scale, originscale; for (blends = 1;blends < MAX_FRAMEBLENDS && frameblend[blends].lerp > 0;blends++) { const short * RESTRICT pose7s = model->data_poses7s + 7 * (frameblend[blends].subframe * model->num_bones + i); @@ -86,33 +136,81 @@ void Mod_Skeletal_AnimateVertices_SSE(const dp_model_t * RESTRICT model, const f dw += -tx*qx - ty*qy - tz*qz; } scale = 1.0f / (rx*rx + ry*ry + rz*rz + rw*rw); + originscale = -model->num_posescale * scale; m[0] = scale*(rw*rw + rx*rx - ry*ry - rz*rz); m[1] = 2*scale*(rx*ry - rw*rz); m[2] = 2*scale*(rx*rz + rw*ry); - m[3] = originscale*scale*(dw*rx - dx*rw + dy*rz - dz*ry); + m[3] = originscale*(dw*rx - dx*rw + dy*rz - dz*ry); m[4] = 2*scale*(rx*ry + rw*rz); m[5] = scale*(rw*rw + ry*ry - rx*rx - rz*rz); m[6] = 2*scale*(ry*rz - rw*rx); - m[7] = originscale*scale*(dw*ry - dx*rz - dy*rw + dz*rx); + m[7] = originscale*(dw*ry - dx*rz - dy*rw + dz*rx); m[8] = 2*scale*(rx*rz - rw*ry); m[9] = 2*scale*(ry*rz + rw*rx); m[10] = scale*(rw*rw + rz*rz - rx*rx - ry*ry); - m[11] = originscale*scale*(dw*rz + dx*ry - dy*rx - dz*rw); + m[11] = originscale*(dw*rz + dx*ry - dy*rx - dz*rw); if (i == r_skeletal_debugbone.integer) m[r_skeletal_debugbonecomponent.integer % 12] += r_skeletal_debugbonevalue.value; m[3] *= r_skeletal_debugtranslatex.value; m[7] *= r_skeletal_debugtranslatey.value; m[11] *= r_skeletal_debugtranslatez.value; - Matrix4x4_FromArray12FloatD3D(&mm, m); - if (model->data_bones[i].parent >= 0) - Matrix4x4_Concat(&bonepose[i], &bonepose[model->data_bones[i].parent], &mm); - else - memcpy(&bonepose[i], &mm, sizeof(mm)); - // create a relative deformation matrix to describe displacement - // from the base mesh, which is used by the actual weighting - Matrix4x4_FromArray12FloatD3D(&mm, model->data_baseboneposeinverse + i * 12); // baseboneposeinverse is 4x3 row-major - Matrix4x4_Concat(&mm2, &bonepose[i], &mm); - Matrix4x4_Transpose(&boneposerelative[i], &mm2); // TODO: Eliminate this transpose + { + const float * RESTRICT n = model->data_baseboneposeinverse + i * 12; + matrix4x4_t * RESTRICT b = &bonepose[i]; + matrix4x4_t * RESTRICT r = &boneposerelative[i]; + __m128 b0, b1, b2, b3, r0, r1, r2, r3, nr; + if (model->data_bones[i].parent >= 0) + { + const matrix4x4_t * RESTRICT p = &bonepose[model->data_bones[i].parent]; + __m128 pr = _mm_load_ps(p->m[0]); + b0 = _mm_mul_ps(pr, _mm_set1_ps(m[0])); + b1 = _mm_mul_ps(pr, _mm_set1_ps(m[1])); + b2 = _mm_mul_ps(pr, _mm_set1_ps(m[2])); + b3 = _mm_mul_ps(pr, _mm_set1_ps(m[3])); + pr = _mm_load_ps(p->m[1]); + b0 = _mm_add_ps(b0, _mm_mul_ps(pr, _mm_set1_ps(m[4]))); + b1 = _mm_add_ps(b1, _mm_mul_ps(pr, _mm_set1_ps(m[5]))); + b2 = _mm_add_ps(b2, _mm_mul_ps(pr, _mm_set1_ps(m[6]))); + b3 = _mm_add_ps(b3, _mm_mul_ps(pr, _mm_set1_ps(m[7]))); + pr = _mm_load_ps(p->m[2]); + b0 = _mm_add_ps(b0, _mm_mul_ps(pr, _mm_set1_ps(m[8]))); + b1 = _mm_add_ps(b1, _mm_mul_ps(pr, _mm_set1_ps(m[9]))); + b2 = _mm_add_ps(b2, _mm_mul_ps(pr, _mm_set1_ps(m[10]))); + b3 = _mm_add_ps(b3, _mm_mul_ps(pr, _mm_set1_ps(m[11]))); + b3 = _mm_add_ps(b3, _mm_load_ps(p->m[3])); + } + else + { + b0 = _mm_setr_ps(m[0], m[4], m[8], 0.0f); + b1 = _mm_setr_ps(m[1], m[5], m[9], 0.0f); + b2 = _mm_setr_ps(m[2], m[6], m[10], 0.0f); + b3 = _mm_setr_ps(m[3], m[7], m[11], 1.0f); + } + _mm_store_ps(b->m[0], b0); + _mm_store_ps(b->m[1], b1); + _mm_store_ps(b->m[2], b2); + _mm_store_ps(b->m[3], b3); + nr = _mm_loadu_ps(n); + r0 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0))); + r1 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1))); + r2 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2))); + r3 = _mm_mul_ps(b0, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3))); + nr = _mm_loadu_ps(n+4); + r0 = _mm_add_ps(r0, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0)))); + r1 = _mm_add_ps(r1, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1)))); + r2 = _mm_add_ps(r2, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2)))); + r3 = _mm_add_ps(r3, _mm_mul_ps(b1, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3)))); + nr = _mm_loadu_ps(n+8); + r0 = _mm_add_ps(r0, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(0, 0, 0, 0)))); + r1 = _mm_add_ps(r1, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(1, 1, 1, 1)))); + r2 = _mm_add_ps(r2, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(2, 2, 2, 2)))); + r3 = _mm_add_ps(r3, _mm_mul_ps(b2, _mm_shuffle_ps(nr, nr, _MM_SHUFFLE(3, 3, 3, 3)))); + r3 = _mm_add_ps(r3, b3); + _mm_store_ps(r->m[0], r0); + _mm_store_ps(r->m[1], r1); + _mm_store_ps(r->m[2], r2); + _mm_store_ps(r->m[3], r3); + } } } -- 2.39.2